Sheri Zhang | 23adc4c | 2021-01-05 12:48:45 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/Helpers.h" |
| 25 | #include "arm_compute/core/ITensorPack.h" |
| 26 | #include "arm_compute/core/Window.h" |
| 27 | #include "src/core/NEON/NEMath.h" |
| 28 | #include "src/core/NEON/wrapper/wrapper.h" |
Sheri Zhang | 23adc4c | 2021-01-05 12:48:45 +0000 | [diff] [blame] | 29 | #include "src/core/helpers/ScaleHelpers.h" |
| 30 | #include "src/core/utils/ScaleUtils.h" |
| 31 | #include "support/Rounding.h" |
| 32 | |
| 33 | #include <cmath> |
| 34 | #include <cstddef> |
| 35 | |
| 36 | #if defined(__ARM_FEATURE_SVE) |
| 37 | #include <arm_sve.h> |
| 38 | |
| 39 | namespace arm_compute |
| 40 | { |
| 41 | namespace |
| 42 | { |
| 43 | void fp16_sve_scale_nearest(const ITensor *src, ITensor *dst, const ITensor *offsets, |
| 44 | float sampling_offset, bool align_corners, const Window &window) |
| 45 | { |
| 46 | const size_t in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; |
| 47 | const size_t in_stride_w = src->info()->dimension(1) + src->info()->padding().top + src->info()->padding().bottom; |
| 48 | const size_t in_stride_wc = in_stride_w * in_stride_c; |
| 49 | const size_t in_dim_h = src->info()->dimension(2); |
| 50 | |
| 51 | // Compute the ratio between source height and destination height |
| 52 | const auto hr = scale_utils::calculate_resize_ratio(in_dim_h, dst->info()->dimension(2), align_corners); |
| 53 | const auto window_start_x = static_cast<int32_t>(window.x().start()); |
| 54 | const auto window_end_x = static_cast<int32_t>(window.x().end()); |
| 55 | |
| 56 | Window win(window); |
| 57 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 58 | Iterator out(dst, win); |
| 59 | |
| 60 | const uint8_t *in_ptr_start = src->buffer() + src->info()->offset_first_element_in_bytes(); |
| 61 | const unsigned int in_stride_bytes_hwc = src->info()->strides_in_bytes()[3]; |
| 62 | |
| 63 | execute_window_loop(win, [&](const Coordinates & id) |
| 64 | { |
| 65 | const int32_t offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))) * in_stride_c; |
| 66 | const auto in_hi = static_cast<int>(align_corners ? utils::rounding::round_half_away_from_zero((id.z() + sampling_offset) * hr) : std::floor((id.z() + sampling_offset) * hr)); |
| 67 | const int offset_row = in_hi * in_stride_wc; |
| 68 | const auto in_ptr = reinterpret_cast<const float16_t *>(in_ptr_start + in_stride_bytes_hwc * id[3]); |
| 69 | const auto out_ptr = reinterpret_cast<float16_t *>(out.ptr()); |
| 70 | |
| 71 | // Compute S elements per iteration |
| 72 | int x = window_start_x; |
| 73 | svbool_t pg = svwhilelt_b16(x, window_end_x); |
| 74 | do |
| 75 | { |
| 76 | // Store results |
| 77 | svst1_f16(pg, out_ptr + x, svld1_f16(pg, in_ptr + offset + offset_row + x)); |
| 78 | |
| 79 | x += svcntw(); |
| 80 | pg = svwhilelt_b16(x, window_end_x); |
| 81 | } |
| 82 | while(svptest_any(svptrue_b16(), pg)); |
| 83 | }, |
| 84 | out); |
| 85 | } |
| 86 | |
| 87 | void fp16_sve_scale_bilinear(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, |
| 88 | BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, |
| 89 | bool align_corners, const Window &window) |
| 90 | { |
| 91 | // Compute the ratio between source height and destination height |
| 92 | const auto hr = scale_utils::calculate_resize_ratio(src->info()->dimension(2), dst->info()->dimension(2), align_corners); |
| 93 | |
| 94 | Iterator out(dst, window); |
| 95 | const int in_stride_c = src->info()->dimension(0) + src->info()->padding().left + src->info()->padding().right; |
| 96 | const int in_dim_w = src->info()->dimension(1); |
| 97 | const int in_dim_h = src->info()->dimension(2); |
| 98 | const int in_stride_wc = in_stride_c * (in_dim_w + src->info()->padding().top + src->info()->padding().bottom); |
| 99 | |
| 100 | // Don't increment in Y and Z direction for the input tensor |
| 101 | // A pointer to the start of this plane is needed as base for the precomputed offsets |
| 102 | Window win_in(window); |
| 103 | win_in.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 104 | win_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 105 | Iterator in(src, win_in); |
| 106 | |
| 107 | if(border_mode == BorderMode::CONSTANT) |
| 108 | { |
| 109 | using ConstType = typename std::conditional<std::is_same<float16_t, float16_t>::value, half, float16_t>::type; |
| 110 | |
| 111 | const float16_t const_border_value = static_cast<float16_t>(constant_border_value.get<ConstType>()); |
| 112 | execute_window_loop(window, [&](const Coordinates & id) |
| 113 | { |
| 114 | const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 115 | const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 116 | const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 117 | const int32_t in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); |
| 118 | const float16_t *in_ptr = reinterpret_cast<const float16_t *>(in.ptr()) + offset * in_stride_c + in_hi * in_stride_wc; |
| 119 | |
| 120 | const auto a00 = (0 <= offset && offset < in_dim_w && 0 <= in_hi && in_hi < in_dim_h) ? *in_ptr : const_border_value; |
| 121 | const auto a01 = (-1 <= offset && offset < in_dim_w - 1 && 0 <= in_hi && in_hi < in_dim_h) ? *(in_ptr + in_stride_c) : const_border_value; |
| 122 | const auto a10 = (0 <= offset && offset < in_dim_w && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_wc) : const_border_value; |
| 123 | const auto a11 = (-1 <= offset && offset < in_dim_w - 1 && -1 <= in_hi && in_hi < in_dim_h - 1) ? *(in_ptr + in_stride_c + in_stride_wc) : const_border_value; |
| 124 | |
| 125 | *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); |
| 126 | }, |
| 127 | in, out); |
| 128 | } |
| 129 | else if(border_mode == BorderMode::REPLICATE) |
| 130 | { |
| 131 | execute_window_loop(window, [&](const Coordinates & id) |
| 132 | { |
| 133 | const auto offset = *reinterpret_cast<const int32_t *>(offsets->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 134 | const auto dx_val = *reinterpret_cast<const float *>(dx->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 135 | const auto dy_val = *reinterpret_cast<const float *>(dy->ptr_to_element(Coordinates(id.y(), id.z()))); |
| 136 | const int in_hi = std::floor((id.z() + sampling_offset) * hr - sampling_offset); |
| 137 | |
| 138 | auto clamped_w = utility::clamp<int>(offset, 0, in_dim_w - 1); |
| 139 | auto clamped_w1 = utility::clamp<int>(offset + 1, 0, in_dim_w - 1); |
| 140 | auto clamped_h = utility::clamp<int>(in_hi, 0, in_dim_h - 1); |
| 141 | auto clamped_h1 = utility::clamp<int>(in_hi + 1, 0, in_dim_h - 1); |
| 142 | |
| 143 | const auto a00 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h * in_stride_wc); |
| 144 | const auto a01 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h * in_stride_wc); |
| 145 | const auto a10 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w * in_stride_c + clamped_h1 * in_stride_wc); |
| 146 | const auto a11 = *(reinterpret_cast<const float16_t *>(in.ptr()) + clamped_w1 * in_stride_c + clamped_h1 * in_stride_wc); |
| 147 | |
| 148 | *reinterpret_cast<float16_t *>(out.ptr()) = static_cast<float16_t>(scale_helpers::delta_bilinear(a00, a01, a10, a11, dx_val, dy_val)); |
| 149 | }, |
| 150 | in, out); |
| 151 | } |
| 152 | else |
| 153 | { |
| 154 | ARM_COMPUTE_ERROR("Not implemented"); |
| 155 | } |
| 156 | } |
| 157 | } |
| 158 | namespace cpu |
| 159 | { |
| 160 | void fp16_sve_scale(const ITensor *src, ITensor *dst, const ITensor *offsets, const ITensor *dx, const ITensor *dy, |
| 161 | InterpolationPolicy policy, BorderMode border_mode, PixelValue constant_border_value, float sampling_offset, |
| 162 | bool align_corners, const Window &window) |
| 163 | { |
| 164 | if(policy == InterpolationPolicy::BILINEAR) |
| 165 | { |
| 166 | fp16_sve_scale_bilinear(src, dst, offsets, dx, dy, border_mode, constant_border_value, sampling_offset, align_corners, window); |
| 167 | } |
| 168 | else if(policy == InterpolationPolicy::NEAREST_NEIGHBOR) |
| 169 | { |
| 170 | fp16_sve_scale_nearest(src, dst, offsets, sampling_offset, align_corners, window); |
| 171 | } |
| 172 | } |
| 173 | } // namespace cpu |
| 174 | } // namespace arm_compute |
| 175 | |
| 176 | #endif // __ARM_FEATURE_SVE |